Chips, Defectors, and IPOs: The Three-Front War for AI Dominance

Chips, Defectors, and IPOs: The Three-Front War for AI Dominance

OpenAI just built its own processor, Google's best researchers keep walking out, and two AI giants are racing each other to Wall Street — all in the same week.

OpenAI Just Got Out From Under Nvidia's Shadow

For years, OpenAI ran its entire empire on borrowed silicon. It leased GPU clusters from Microsoft, paid Nvidia's eye-watering prices for H100s and H200s, and treated compute as someone else's problem. That era ended June 24, when OpenAI and Broadcom jointly unveiled Jalapeño — a custom inference chip designed from scratch for the specific demands of large language models.

The name is playful. The ambition is not. According to Broadcom CEO Hock Tan, Jalapeño cuts inference costs by roughly 50% compared to standard AI GPUs. For a company that burns through hundreds of millions of dollars annually running ChatGPT queries, that's not an efficiency gain — it's existential arithmetic. If the chip delivers on its projections when it rolls out to production data centers later this year, OpenAI will have pulled off something that took Google nearly a decade: owning the full stack from model weights to the silicon they run on.

What makes Jalapeño's origin story almost more impressive than the chip itself is the timeline. The accelerator went from design brief to manufacturing tape-out in nine months — what Broadcom describes as the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors. And here's the recursive twist: OpenAI's own models helped design it, running simulations that would have taken traditional chip engineers months to complete manually.

The chip is purpose-built for inference (running trained models in response to user inputs), not for training. That distinction matters. Training a frontier model still requires Nvidia's H100 clusters or similar brute-force GPU arrays. But inference is where the real cost pressure lives — every ChatGPT conversation, every API call from a developer, every enterprise automation that runs thousands of queries per hour. Owning that compute layer is a strategic moat of the highest order.

Microsoft, meanwhile, is moving in a parallel direction. At its Build 2026 conference earlier this month, the company launched seven in-house AI models under the MAI brand, including MAI-Thinking-1, its first proprietary reasoning model trained entirely without OpenAI data. The message was diplomatic but unmistakable: the most important financial backer of OpenAI is quietly building contingency plans, co-designing its MAI models with its own Maia 200 silicon and reporting a 1.4x performance-per-watt gain in the process.

The silicon arms race is no longer a subplot. It's the main event.

Google Is Watching Its Future Walk Out the Door

Jonas Adler and Alexander Pritzel are not household names outside of AI research circles. Inside Google DeepMind, they were load-bearing walls. Adler worked on the company's AI coding efforts; Pritzel was deeply involved in training methodologies for the Gemini model family. Both are leaving for Anthropic, Bloomberg reported June 24.

They join a parade that is becoming harder for Google to wave off as normal attrition. Nobel Prize winner John Jumper left for Anthropic earlier this year. Noam Shazeer — one of the original authors of the "Attention Is All You Need" paper that sparked the entire transformer era — decamped to OpenAI. Senior research engineer Arthur Conmy announced his own Anthropic move, specifically to work on AI safety.

The economic pull is obvious: both Anthropic and OpenAI have filed confidentially for IPOs, and researchers joining now stand to receive equity that could be worth multiples of whatever Google stock they're leaving behind. TechCrunch noted the dynamic succinctly — two startups on the cusp of going public are offering something a Big Tech salary can't easily match: the chance at a rare, pre-IPO payday.

But the talent drain carries strategic weight beyond individual departures. Google's Gemini models have consistently ranked among the world's top frontier systems. The researchers building and refining them have accumulated institutional knowledge — architecture intuitions, training recipe expertise, the kind of hard-won insight that doesn't transfer via documentation. When those people move, some of that knowledge moves with them. Google isn't running dry; it still employs thousands of world-class researchers. But each high-profile exit sends a signal to the researchers still on the fence.

Two AI Giants, One Wall Street Finish Line

The IPO race crystallized in early June. Anthropic filed confidentially for a public listing around June 1; OpenAI followed roughly a week later. Both companies believe that whoever lists first shapes how institutional investors benchmark the entire AI sector.

The competition has teeth. Anthropic now commands more than 40% of the generative AI coding market, compared to OpenAI's 21%, according to recent market analysis. OpenAI retains the consumer mindshare advantage — ChatGPT remains the default entry point for hundreds of millions of users — but Anthropic's penetration in enterprise developer workflows has become a serious revenue story. Claude's reputation for longer context windows and lower hallucination rates has made it the preferred engine for coding assistants, legal document review, and research automation.

The IPO filings arrive at a complicated moment. Both companies are watching their unit economics tighten. Enterprise customers who once threw compute budgets at every AI use case are now auditing which workloads actually generate ROI. The era of "tokenmaxxing" — running massive prompts through frontier models just because you can — is giving way to a more sober, cost-sensitive discipline. The pressure is visible in pricing: both labs have been steadily reducing per-token costs as competition from cheaper alternatives, including Chinese models like DeepSeek and Alibaba's Qwen family, forces their hand.

OpenAI is seeking a valuation of around $1 trillion. Anthropic hasn't disclosed a target. But both numbers will be scrutinized through one lens above all others: can these companies build durable businesses, or are they perpetually burning investor capital to train the next model?

Europe Loosens Its Grip — For Now

Across the Atlantic, the European Union's AI Act took a quieter turn in mid-June. The European Parliament granted final approval to a package of amendments on June 16 that delay certain significant obligations for high-risk AI systems and simplify several regulatory requirements. The amended rules are expected to be formally in effect by August 2.

The delays are a partial concession to industry criticism that the original timelines were unworkable. High-risk AI systems — tools used in hiring, credit scoring, law enforcement, and medical diagnosis — now have more runway before full compliance kicks in. European tech companies had argued that the original Act threatened to put EU-based AI developers at a permanent disadvantage against American and Chinese competitors operating under looser rules.

The amendments don't gut the Act's ambitions; transparency requirements and prohibitions on certain AI uses remain intact. But they represent a recalibration: Brussels is still committed to the most comprehensive AI governance framework in the world, just not at a pace that would strangle the companies trying to build within it.

The Bottom Line

The week's stories share a single through-line: AI is no longer primarily a software story. It's a story about who controls the infrastructure — the chips, the talent pipelines, the capital markets access, and the regulatory environment those systems operate within. OpenAI designing its own silicon, Google hemorrhaging researchers to IPO-bound rivals, Anthropic and OpenAI racing to Wall Street, the EU deciding how hard to clamp down — each arc is separate, but they're all pulling on the same rope. Whoever commands the full stack by 2027 won't just win the current AI race. They'll set the terms for the next one.

Sources

• OpenAI and Broadcom Unveil LLM-Optimized Inference Chip — OpenAI
• OpenAI unveils its first custom chip, built by Broadcom — TechCrunch
• OpenAI unveils first custom AI inference chip, Jalapeño, with Broadcom — VentureBeat
• AI researchers continue to leave Google for its rivals — TechCrunch
• Google Poised to Lose Two More High-Profile AI Staffers to Anthropic — Bloomberg
• It's not about Anthropic vs. OpenAI anymore — TechCrunch
• OpenAI and Anthropic face new AI reality as users shift from 'tokenmaxxing' to efficiency — CNBC
• EU Approves Delays and Other Amendments to Certain EU AI Act Obligations — Morgan Lewis
• Building a hill-climbing machine: Launching seven new MAI models — Microsoft AI
• Microsoft unveils new AI models to lessen reliance on OpenAI — CNBC